Introducing surgical workflow recognition in orthopaedic surgery with timestamp supervision.

Area of Science:

• Computer Vision
• Medical Informatics
• Orthopaedic Surgery

Background:

• Surgical workflow recognition (SWR) is crucial for improving patient safety and surgeon performance.
• Existing SWR research primarily focuses on endoscopic procedures due to limited open surgery datasets.
• Open surgery presents unique challenges like occlusions and varied camera angles for SWR.

Purpose of the Study:

• To develop and validate SWR for open orthopaedic surgery, specifically minimally invasive plate osteosynthesis (MIPO) for distal radius fractures (DRFs).
• To introduce the largest publicly available dataset for open surgery SWR.
• To propose a novel weakly-supervised method to reduce annotation burden in SWR.

Main Methods:

• Creation of a new dataset comprising 50 videos (approx. 30 hours) of DRF MIPO procedures.
• Evaluation of state-of-the-art fully-supervised methods to establish a baseline for MIPO surgical step prediction.
• Development of the Uncertainty- and Cluster-Aware Temporal Diffusion (UCATD) method for weakly-supervised SWR using timestamp annotations.

Main Results:

• The introduced dataset is the largest publicly available for open surgery SWR.
• UCATD significantly outperforms existing timestamp-supervised methods.
• UCATD achieves competitive results with fully-supervised methods using only 0.1% annotated data.

Conclusions:

• The developed dataset and UCATD method advance SWR in open orthopaedic surgery.
• Weakly-supervised learning, particularly UCATD, offers a viable solution to the annotation challenge in SWR.
• This work paves the way for improved SWR applications in complex open surgical procedures.